The point, without the rewrite.

What it does

Paste an article, a report, a set of meeting notes, or a long transcript. The tool breaks the text into sentences, measures which content words repeat, and gives each sentence a score. It then keeps a requested number of sentences — either a percentage of the source or a fixed count — and puts the winners back in the order they appeared.

That is the whole operation. There is no hidden prompt, no model provider, and no assistant writing a new paragraph. The output is a shorter version of your own text, assembled from sentences you can find in the input. If a sentence is not in the source, it cannot appear in the summary.

The workbench also shows the topic words that influenced the ranking and a short “why these sentences” list with a score for each kept sentence. Those details matter: a summary is easier to trust when you can see what the selector was optimising for. They are not a claim that the ranking understands the argument.

How sentences are ranked

The scorer starts by lowercasing the text and splitting it into word tokens. It removes a built-in list of common English stopwords — words such as “the”, “and”, “of”, and “that” — and applies a light stemming step so that simple variants such as “report” and “reports” count as the same term. It then counts how often each remaining term appears across the whole text.

A sentence receives credit for containing frequent content words, but the score is divided by sentence length so a long sentence cannot win merely by having more words. A small position bonus favours the opening sentences, and very short or extremely long sentences are damped. After ranking, the tool compares candidate sentences with the ones already selected. If two sentences are too similar, the later one is skipped so the summary does not spend its budget repeating the same point.

This is a frequency-and-position ranker, not a neural summarizer and not a full TextRank implementation. It is deterministic: the same text and settings produce the same result. That predictability is useful when you want to compare two drafts, but it also means the tool has no idea which claim is true, which caveat matters, or which sentence is the actual thesis.

Choosing a length

The length control is a request, not a guarantee of quality. A percentage is useful when the source is long and you want a predictable ratio: 20–30% is a practical starting range for an article or report. Five to eight sentences is a reasonable target for a one-page brief or a quick read of meeting notes. For a very long document, 10% may still produce a useful triage summary.

Start short. Read the result against the source and ask a simple question: is the sentence that carries the main point present? If not, raise the length or choose the sentence-count mode and add two or three sentences. A summary that contains every minor detail is not a summary; it is the document with extra steps.

Short inputs are handled conservatively. If the source has only two sentences, both are kept. If the requested count is larger than the number of sentences, the tool returns all of them rather than padding the output. When near-duplicate sentences are skipped, a second pass fills the target from the remaining highest-ranked sentences, so the count is honoured whenever the source has enough distinct material.

Extractive, not generative

Generative summarizers can smooth a source into a new paragraph. They can merge two ideas, replace a clumsy phrase, and add a transition that was never there. They can also invent a number, reverse a caveat, or confidently state something the source does not say. This tool makes the opposite trade. It is limited, literal, and auditable.

The practical consequence is that the output may sound uneven. A selected sentence can begin with “However” even though the sentence it referred to was not selected. A pronoun can point at a noun that stayed behind. The summary is a set of signposts back to the source, not a replacement for it. For a legal document, a medical note, or a financial report, that limitation is a feature: every sentence can be checked in context.

Use the result for triage, for a first pass over a long document, or for extracting candidate sentences that you will edit by hand. Do not present it as an abstract you have not read against the original.

Examples

The pilot began in March with two support teams. Average first-response time fell from nine minutes to six. The team handled 18 percent more tickets without adding headcount. The improvement was largest during the Monday morning peak. The pilot cost less than expected because it reused the existing dashboard. The operations lead recommended extending it to the billing team in July.

The pilot began in March with two support teams. The team handled 18 percent more tickets without adding headcount. The operations lead recommended extending it to the billing team in July.

The selected sentences cover the setup, the measurable result, and the recommendation. The sentence about the Monday peak and the sentence about cost are plausible details, but they are not the spine of the update. A different length setting could include them.

A good changelog answers three questions. What changed? Who does it affect? What should the reader do next? Most changelogs answer the first question and stop. They list merged pull requests and internal ticket numbers, which are useful to the team that shipped the work but not to the person using the product. A better entry names the user-visible change, states whether it is a fix or a new behaviour, and links to the detail only when the reader needs it.

A good changelog answers three questions. What changed? Who does it affect? What should the reader do next? A better entry names the user-visible change, states whether it is a fix or a new behaviour, and links to the detail only when the reader needs it.

The three short questions survive because “changelog”, “changed”, and “reader” recur in the paragraph and the questions are not long enough to be penalised. This is a useful reminder that frequency scoring rewards repetition, not rhetorical importance. If the questions were buried in the middle and mentioned once, the result would be different.

Limits worth knowing

• Sentence splitting is heuristic. The splitter looks for terminal punctuation and trims each sentence. Abbreviations, decimals, ellipses, and missing punctuation can create a bad boundary, and no splitter is perfect.
• Frequency is not understanding. A sentence that states the main point once may lose to a sentence that repeats a side issue four times. Negation is especially fragile: “the policy does not apply” and “the policy applies” share most of their words.
• Stopwords are English. In another language, the keyword table may still contain common grammatical words, and the ranking will be weaker. The punctuation and sentence rules are also oriented toward English prose.
• The source is capped at 200,000 characters. That is generous for an article or a report, but it is not a book-length tool. A very large paste can also make the browser work hard; shorten the source or work in sections.
• Nothing is stored. The text lives in the page while it is open. Copy the summary before you leave, because a refresh clears both the input and the output.

Page updated 27 October 2026